Approximate Bayesian Computation via Sufficient Dimension Reduction

TL;DR

This paper advances Approximate Bayesian Computation by establishing its theoretical properties and introducing a semi-automatic algorithm leveraging sufficient dimension reduction to improve efficiency in complex models.

Contribution

It provides theoretical insights into ABC's limiting behavior and proposes a novel SDR-based semi-automatic algorithm for more efficient Bayesian inference.

Findings

Proved theorems on ABC's limiting behavior.

Developed a new SDR-based ABC algorithm.

Demonstrated connections between SDR and ABC.

Abstract

Approximate Bayesian computation (ABC) has gained popularity in recent years owing to its easy implementation, nice interpretation and good performance. Its advantages are more visible when one encounters complex models where maximum likelihood estimation as well as Bayesian analysis via Markov chain Monte Carlo demand prohibitively large amount of time. This paper examines properties of ABC both from a theoretical as well as from a computational point of view.We consolidate the ABC theory by proving theorems related to its limiting behaviour. In particular, we consider partial posteriors, which serve as the first step towards approximating the full posteriors. Also, a new semi-automatic algorithm of ABC is proposed using sufficient dimension reduction (SDR) method. SDR has primarily surfaced in the frequentist literature. But we have demonstrated in this paper that it has connections with ABC as well.